Skip to content

uniref50-afdb-holdout-filtered 1.0.0 - #24

Merged
briney merged 2 commits into
mainfrom
flyloft/2f2cbf3a
Nov 21, 2026
Merged

briney merged 2 commits into
mainfrom
flyloft/2f2cbf3a

Conversation

@briney

@briney briney commented Nov 21, 2026

Copy link
Copy Markdown
Member

Cook uniref50-afdb-holdout-filtered 1.0.0: build resumes from the staged MMseqs2 filter handed off at d8537b0, then test and push.

Cooked by flyloft campaign 2f2cbf3a; its record is results/uniref50-afdb-holdout-filtered/1.0.0.txt.

…e homolog of an afdb-rep-stok-large holdout protein

Every UniRef50 entry with an MMseqs2 hit at >=50% sequence identity and
>=80% coverage on both sides (--min-seq-id 0.5 -c 0.8 --cov-mode 0 -s 7.5
-e 100 --max-seqs 2000, every alignment re-checked in code) against any
afdb-rep-stok-large@v1.0.1 validation or test protein is removed, so
sequence pretraining cannot leak the structure dataset's held-out split:
26,523 of 41,243,524 entries removed, 41,217,001 survive in uniref50's own
layout, order and 1M-row zstd shards. The recipe chains the two published
datasets, stages its work resumably under <out>/.work, and its test
re-searches sampled removed ids under the filter's own flags.
@briney
briney marked this pull request as ready for review November 21, 2026 14:34
@briney
briney merged commit 4f9f06f into main Nov 21, 2026
1 check passed
@briney
briney deleted the flyloft/2f2cbf3a branch November 21, 2026 14:35
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant